« first day (1918 days earlier)      last day (3146 days later) » 
00:00 - 20:0020:00 - 00:00

8:00 PM
meh, if they are good edits let the review queue flood for all I care...
@ryanyuyu To be fair, there are only 16 edits and those were over a 20-30 minute period of time. There are human reviewers/editors that do more than that.
That's a good point, but the humans who do that also annoy me.
Current accuracy: 41.17%. 60% has been suggested as a target.
yes, I know
8:03 PM
SO has never had issues clearing the suggested edit queue. Until that happens, the 'don't flood' argument is moot IMO
@ryanyuyu SO's average approval rate is 80.43%. I'd be doing incredibly well if a machine can achieve that, so I think it's probably better to start with something lower. Once it's doing a majority i.e. over 50% good edits, it's definitely worth something and we can start looking at hitting that 80.
Well ideally the bot does better than most users...
As a proof of concept or other milestone goal I suppose %60 is ok. But the final goal should be higher.
@ryanyuyu Ideally. But we're working in the field of computer science and artificial intelligence, where that doesn't happen in the majority of projects. We target being worth something first, then we target achieving average, then we target being better.
Intermediate value theorem. To hit 100%, you've gotta hit 60%.
@ryanyuyu Absolutely. I want the final goal to be 100%. But setting that as a target right now is unrealistic, so we go for something lower.
8:06 PM
There are +160000 available review tasks everyday
Ok what does the %60 goal represent?
You may have an order of magnitude error there.
Then again, maybe not. shrug
@ryanyuyu A way to get to 100%
@ryanyuyu An intermediate target. 60% was a suggestion from @EricLeschinski - I think of it as a point at which the bot is doing clearly useful work.
100% good edits, or 100% approved edits? There's a difference....
8:07 PM
and we come again to the input information: how we define a "good edit"
Or, to be precise - how do we define a good retag .
oh, that's easy
We know it when we see it
8:09 PM
one that allows an unanwered question to reach the people that can answer it
^ that
That's actually a really good metric.
Not objective, though
that should be the only metric IMHO
I think it could be objective
8:10 PM
@S.L.Barth good. Now that some have been rejected, I will technically never reach 100% approved.
@Undo actually, you can demostrate it
@ArtOfCode Sliding window, of course
@Undo Of course, but in technicality ;)
@Braiam You can try, but you really can't prove 100%
look for an unanswered question with a bad tag combination, retag, wait for it for be resolved
8:11 PM
@ArtOfCode I've also seen good edits get rejected. So "100% approved" would be a bad yardstick.
Is there a way to get even just compare view counts. That's fairly naive, but could be useful.
where resolved means: answered or closed
@Braiam closed as a duplicate, or closed in general?
Right. Suggestions, please. What should the algorithm be?
Would just "voted on", mean it reached someone?
8:12 PM
@S.L.Barth general, the people that can answer it knows what it lacks to be answered too
@ArtOfCode Don't need one yet
There's API docs if you want to look at them
@Undo The bot's, I mean. How should I determine what tags to add?
No idea
@Andy I'd prefer a number of votes. Or a number of upvotes.
@ArtOfCode You know you've visited that URL frequently when you only have to type ap<enter> and it opens the docs
8:13 PM
@Andy I have that :)
I have ideas on what tags it shouldn't add, but that needs access to deleted questions
@S.L.Barth Downvotes are valid too.
We've had the suggestion to avoid turd-polishing that we only act on questions scored >= {threshold}, which is useful. I'll put that in.
That's definitely something important.
There's also been a suggestion to check for "DO NOT USE THIS TAG" in the wiki, though I'm not sure how useful that is - not all bad tags have it, and it uses API quota.
8:15 PM
a cache of that can help
you could potentially use a SEDE query to hardcode certain ignored tags.
A quota of 10K a day. How many edits are you going to make and do you plan on doing something different when the bot inevitably reaches 2K rep?
Ooh nefarious idea:
Donate rep as bounty
kinda sketchy
I thought you couldn't get more than 1K from edits?
oh that's true
8:17 PM
You're right @S.L.Barth. As long as the bot doesn't do anything else, it'll always need to have it's edits approved
@ArtOfCode other SEDE query might exist, but here's a starting point for determining your bot's accuracy once the data makes it to SEDE. Feel free to use/edit/ignore it. data.stackexchange.com/stackoverflow/query/385100/…
@ArtOfCode You might also want to keep track of tags that have been nominated for burnination. To save the API quota, you could just store these in a database or cache (as @Braiam suggests for the "DO NOT USE" tags).
@Andy You can gain a maximum of 1000 rep from suggested edits alone. Without posting anything, the bot will never reach 2k.
oh, someone beat me to that
@ryanyuyu Bad plan. You'd get done for it.
oh yeah I'm sure. Actually very sure.
So how do I improve this thing beyond that?
8:26 PM
@ArtOfCode Beyond 1K ? I wouldn't worry about that for now.
@S.L.Barth No, beyond the ideas I've already got. Not worried about rep.
I'm thinking a "blacklist" of tags that should not be added.
I'm not sure. I personally skip most tag-only suggestions because I can't judge them accurately. Unless I'm active in those tags, I really don't know.
You could look into code sections in posts, and see if calls from certain API's are used.
8:45 PM
[ SmokeDetector ] Few unique characters in answer: Android : Unable to display multiple lines of text in Notification by Black Nigga on stackoverflow.com
@SmokeDetector tpu-
9:42 PM
[ SmokeDetector ] Bad keyword in title: Can Wolverine build muscle mass? by phantom42 on scifi.stackexchange.com
@Undo there's a small benefit plus the downside of wasting peoples' time in review and with bumping.
Not worth it, no significant change, only an annoyance.
@bjb568 Bumping on SO means nothing, and this is by far not the worst thing people volunteering in the suggested edit review queue see.
But it's not approve-worthy.
If you don't think it is, don't approve them
It directly violates the editing help, and I honestly would be fine with using them as known-bad audits.
9:45 PM
If enough people agree with you, it will naturally discourage this.
But this is not something to address directly through a "no bots" policy
And tag-only edits do improve something, otherwise we shouldn't have tags.
And heck no to using them as audits
This is behavior that I've flagged several times (on humans) for abuse. It's gaming the broken roboified edit review system.
@Undo not substantially
@bjb568 not substantially != "none whatsoever"
Suggested edits must improve multiple aspects of a post. Trivial edits are explicitly and strongly discouraged.
> This edit does not make the post even a little bit easier to read, easier to find, more accurate or more accessible. Changes are either completely superfluous or actively harm readability.
@Undo it's none when you factor in the harm: it's noise
9:48 PM
There's a reason we got rid of the 'too minor' reason
@bjb568 On SO, there is no noise
Everything is noise, so nothing is noise
3 hours ago, by Undo
I used to like making those kind of ideological arguments, but then I realized we're trying to run a real site, in real life
@Undo so fine, write a custom reject reason if you want, but most rejectors would say that captures the spirit
eh, I'd be fairly ticked off at those reviewers.
You can use autoreview comments or something to rubber-stamp "don't edit crap" and "your edit sucks and does nothing" if you want.
10:48 PM
[ SmokeDetector ] Offensive answer detected: Why do I keep losing weapons in GTA V online when I save? by Veselin on gaming.stackexchange.com
[ SmokeDetector ] Few unique characters in body, repeating characters in body: Problems with this C code by Pepe Pérez on stackoverflow.com
[ SmokeDetector ] Bad keyword in body: Retrieve browser history on samsung smart tv? by Cmackattack77 on stackoverflow.com
11:05 PM
[ SmokeDetector ] Pattern-matching website in title, url in title: http://healthsupplementcare.com/cognimaxx-xl/ by sakyu juriya on drupal.stackexchange.com
need a final close vote here: stackoverflow.com/questions/33533631/…
@hichris123 tnx
@SmokeDetector NAA
@SmokeDetector tpu- gone
11:51 PM
sd why
@misterManSam [:4154337] Title - Position 21-35: support number
sd - ignore
1. [:4154349] <skipped>
2. [:4154337] Post ignored; alerts about it will no longer be posted.
00:00 - 20:0020:00 - 00:00

« first day (1918 days earlier)      last day (3146 days later) »